pulp::gpu_audio::GpuWaveNetSession Class Reference

A prepared, one-stream shared-memory WaveNet session. More...

#include <gpu_wavenet.hpp>

Classes
struct	Config
struct	CreateResult

Public Member Functions
	~GpuWaveNetSession ()
	GpuWaveNetSession (const GpuWaveNetSession &)=delete
GpuWaveNetSession &	operator= (const GpuWaveNetSession &)=delete
bool	prepared () const noexcept
std::uint32_t	block_size () const noexcept
GpuWaveNetCompletionPolicy	completion_policy () const noexcept
	Returns the requested policy.
bool	completion_policy_supported () const noexcept
	True when the provider accepted the requested policy without fallback.
bool	submit_block (std::span< const float > input, std::uint64_t sequence, std::uint64_t deadline_ns=0) noexcept
	Copies one mono block into a persistent shared slot and submits it to the authenticated provider.
std::size_t	service (std::uint64_t now_ns) noexcept
	Services already-submitted provider work without waiting for a future completion.
std::size_t	service_until (std::uint64_t deadline_ns) noexcept
	Services completions until the supplied steady-clock deadline.
std::optional< GpuWaveNetBlockResult >	receive (std::span< float > output) noexcept
	Copies one completed block into output and releases its shared slot.
bool	release () noexcept
	Host/quiescent-only release.
bool	release_bounded (std::uint64_t budget_ns) noexcept
	release() that bounds the provider's physical drain to budget_ns.
bool	device_lost () const noexcept
	True once the provider has observed loss of its device.
std::uint64_t	outstanding_blocks () const noexcept
	Blocks accepted by submit_block() whose result receive() has not yet returned, whether delivered or failed.

Static Public Member Functions
static CreateResult	create (const Config &config) noexcept

Friends
struct	detail::GpuWaveNetSessionAccess
struct	detail::GpuWaveNetSessionTestAccess
struct	detail::GpuWaveNetSessionPoolAccess

Detailed Description

A prepared, one-stream shared-memory WaveNet session.

The session owns the provider, resident model resources, causal history, and fixed shared input/output slots. Callers provide only model metadata, weights, and audio spans; Dawn, Metal, queue, and slot handles remain private. submit_block() and service() are serialized non-realtime operations in this first SDK seam. A caller that needs a realtime bridge should connect this session to its own preallocated callback/fallback machinery rather than invoking provider work from the audio callback.

The first public session is intentionally limited to one causal stream (descriptor.stream_instances == 1). Submitted sequence numbers must be contiguous because they advance the persistent WaveNet history.

Constructor & Destructor Documentation

~GpuWaveNetSession()

pulp::gpu_audio::GpuWaveNetSession::~GpuWaveNetSession

(

)

GpuWaveNetSession()

pulp::gpu_audio::GpuWaveNetSession::GpuWaveNetSession ( const GpuWaveNetSession & )

delete

Member Function Documentation

block_size()

std::uint32_t pulp::gpu_audio::GpuWaveNetSession::block_size ( ) const

noexcept

completion_policy()

GpuWaveNetCompletionPolicy pulp::gpu_audio::GpuWaveNetSession::completion_policy ( ) const

noexcept

Returns the requested policy.

It remains observable even when the backend cannot provide that policy and falls back to its default.

completion_policy_supported()

bool pulp::gpu_audio::GpuWaveNetSession::completion_policy_supported ( ) const

noexcept

True when the provider accepted the requested policy without fallback.

create()

CreateResult pulp::gpu_audio::GpuWaveNetSession::create ( const Config & config )

staticnoexcept

device_lost()

bool pulp::gpu_audio::GpuWaveNetSession::device_lost ( ) const

noexcept

True once the provider has observed loss of its device.

A lost session never delivers again; its owner must release it and create a new one. Serialized non-realtime diagnostic, like service().

operator=()

GpuWaveNetSession & pulp::gpu_audio::GpuWaveNetSession::operator= ( const GpuWaveNetSession & )

delete

outstanding_blocks()

std::uint64_t pulp::gpu_audio::GpuWaveNetSession::outstanding_blocks ( ) const

noexcept

Blocks accepted by submit_block() whose result receive() has not yet returned, whether delivered or failed.

prepared()

bool pulp::gpu_audio::GpuWaveNetSession::prepared ( ) const

noexcept

receive()

std::optional< GpuWaveNetBlockResult > pulp::gpu_audio::GpuWaveNetSession::receive ( std::span< float > output )

noexcept

Copies one completed block into output and releases its shared slot.

A failed provider completion returns a result with no output written.

release()

bool pulp::gpu_audio::GpuWaveNetSession::release ( )

noexcept

Host/quiescent-only release.

A false result retains the session so the caller can retry the provider's physical drain barrier.

release_bounded()

bool pulp::gpu_audio::GpuWaveNetSession::release_bounded ( std::uint64_t budget_ns )

noexcept

release() that bounds the provider's physical drain to budget_ns.

On expiry the provider abandons to quarantine: its device is destroyed, its shared memory stays owned, and this returns false. The session keeps ownership on false; the caller must still call release() later to complete the disposal. Host/quiescent or serialized worker.

service()

std::size_t pulp::gpu_audio::GpuWaveNetSession::service ( std::uint64_t now_ns )

noexcept

Services already-submitted provider work without waiting for a future completion.

Returns the number of newly visible terminal records.

service_until()

std::size_t pulp::gpu_audio::GpuWaveNetSession::service_until ( std::uint64_t deadline_ns )

noexcept

Services completions until the supplied steady-clock deadline.

This is serialized non-realtime work; never call it from an audio callback.

submit_block()

bool pulp::gpu_audio::GpuWaveNetSession::submit_block ( std::span< const float > input, std::uint64_t sequence, std::uint64_t deadline_ns = 0 )

noexcept

Copies one mono block into a persistent shared slot and submits it to the authenticated provider.

deadline_ns == 0 disables late marking.

Friends And Related Symbol Documentation

detail::GpuWaveNetSessionAccess

friend struct detail::GpuWaveNetSessionAccess

friend

detail::GpuWaveNetSessionPoolAccess

friend struct detail::GpuWaveNetSessionPoolAccess

friend

detail::GpuWaveNetSessionTestAccess

friend struct detail::GpuWaveNetSessionTestAccess

friend

---

The documentation for this class was generated from the following file:

• /home/runner/work/pulp/pulp/core/gpu_audio/include/pulp/gpu_audio/gpu_wavenet.hpp